search

wyf0912 / LLFlow

[AAAI 2022] The code release of paper "AAAI Low-Light Image Enhancement with Normalizing Flow"
Other
291 stars 36 forks source link

How to use only Invertible Neural Networks？ #18

Closed kaixindeyuesefen closed 2 years ago

kaixindeyuesefen commented 2 years ago

How to use only Invertible Neural Networks？

wyf0912 commented 2 years ago

Do you mean in the evaluation phase? or the training phase?

kaixindeyuesefen commented 2 years ago

training.

wyf0912 commented 2 years ago

You could try "Glow: Generative Flow with Invertible 1x1 Convolutions". However, if there are no conditional features, it could be quite difficult to get converged on a high-resolution and diverse dataset for a flow-based model at current.

kaixindeyuesefen commented 2 years ago

Thank you very much!

kaixindeyuesefen commented 2 years ago

Hello, I have another question to ask you. When I used your "NoEncoder" to train the model, the picture I got was black. May I ask how I can modify your "NoEncoder" so that I can only use invertible network to train the model? Because I saw that every feature of "result" in your NoEncoder was ×0, I guess this is the reason why it was black.

wyf0912 commented 2 years ago

I originally want to use that encoder for the ablation study. However, I find it's pretty difficult to get converged. So in Table 5 in the paper, I do not report the results that only using the invertible network

kaixindeyuesefen commented 2 years ago

Hello, I recently encountered a little problem when reproducing “Glow” for low-light enhancement. I would like to ask if you still have this part of the code, can you provide it?